That's an interesting idea. I don't know enough to say if using a large offset would be enough to mitigate the downsides of +=2, i would have to read more about it. In actual use you can go directly to using random candidates for each test as a few extra seconds to generate 2 primes would be worth the benefits. In hindsight my simplistic rng() can also probably be better optimized (batch load random bits and cache) to make it faster as the slowdown primarily comes from repeated filesystem access.